The 2040 Pause Nobody's Actually Taking

Someone built a website called Plan A. The idea: talk the US and China into a treaty that freezes frontier AI at roughly top-human-expert level and holds it there until 2040, with full transparency and shared compute rules underneath it, everybody agreeing not to race each other off a cliff.

I want to believe it. I don't.

What Plan A actually asks for

Read past the framing and Plan A isn't really a prediction. Even ai-2040.com admits that up front: it's a recommendation, not a forecast of what's going to happen. That's a bigger admission than it sounds. A recommendation is what you'd like to happen. A forecast is what you think will. Most of the site blurs the line between the two.

The actual ask is a US-China agreement, worldwide verification of AI R&D, several companies scaling in lockstep, and a hard stop at expert-human-level capability that holds for fourteen years. The urgency isn't invented, either. Their own numbers put full automation of AI R&D around 2030 if nothing changes, which isn't a comfortable thing to sit with.

Wanting that outcome and getting it are two different jobs. Here's where the plan runs into trouble.

China isn't signing this

Start with the deal itself. It only works if China signs, and nothing about China's current position says it will.

Stanford's 2026 AI Index put the performance gap between top US and Chinese models at 1,300 points in May 2023. By March 2026 it was down to 39 (Euronews). China's also ahead now on AI research citations, patents, and robotics deployment. A country closing a gap that fast doesn't sign a treaty to freeze it in place. It signs one once it's ahead, if it signs one at all.

Beijing isn't waiting around to be asked. It launched the World AI Cooperation Organization out of Shanghai in July 2026: 29 founding members, Russia and Brazil among them, the US, UK, and EU left off the list. That reads like a country standing up its own governance track, not one hedging on somebody else's.

Chinese officials did sign the Bletchley Declaration, for what it's worth. That's a safety dialogue running alongside the race, not in place of it.

Nobody can verify what they'd be agreeing to

Say China signs anyway. Plan A calls for complete transparency in AI R&D and worldwide verification. Verification of what, exactly?

Nuclear arms control works, more or less, because you can count warheads and send inspectors to look at hardware that doesn't lie. AI capability has no equivalent. Nobody's built a gauge you point at a data center that reads "top human expert level, stop here." Capability gets measured by benchmark, benchmarks get gamed, and the real frontier work happens behind NDAs on compute clusters no outside inspector will ever see.

Without a working verification mechanism, a pacing agreement is theater. Nobody can confirm compliance, or catch a cheat, until three model generations have already shipped. NAI500 made the same point: pacing rules with no way to measure them don't slow anyone down. They just tell you, after the fact, who kept their word.

The deadline accelerates everything it's supposed to slow

The site glosses over what a hard deadline actually does. Set a pace limit for some future year and you don't get a slowdown leading up to it. You get a sprint.

Every lab and every country with a stake now has a countdown clock and a reason to cram as much progress as possible into the runway before the restriction lands. NAI500 calls this the preemptive acceleration paradox: pacing rules don't cool development, they squeeze it into the window before enforcement starts. A rule meant to slow the race can end up doing the opposite of what it says on the label.

The baseline pace isn't helping the case either. Training compute has grown roughly 5x a year since 2020, doubling every five months. Training costs are climbing 3.5x a year. Software efficiency gains of about 3x a year were supposed to take the pressure off, and mostly just fund the next scale-up instead (Epoch AI). That's not a curve that flattens because a website made a good argument.

Even the industry isn't asking for this

The clearest tell is who's actually asking for what. In July 2026, more than 1,200 employees across Anthropic, OpenAI, Google DeepMind, and Meta, Dario Amodei and OpenAI's Jakub Pachocki among them, signed a letter asking the US government to help build the tools for pacing AI development (Fortune).

Look at what they actually asked for. Not a pause. Not a 2040 freeze. Support for developing the technical and governance tools that might eventually make pacing possible, if the moment calls for it. Fortune called the language notably careful and slightly vague, and that's generous. These are the industry's most safety-minded people, and their strongest public ask stops at "build the tools, maybe use them later." Not "stop."

The people closest to the technology have the most reason to be scared of it and the least reason to bluff about what it can do. When their loudest ask falls well short of what Plan A demands, a website isn't going to get a fourteen-year freeze signed between two superpowers.

The honest caveat

None of this makes the underlying worry fake. The 2030 timeline for automated AI R&D isn't invented, and the sandbox-escape incident that helped prompt that 1,200-signature letter actually happened. Something is moving fast, and the people closest to it are nervous. I think that part of Plan A's premise is right.

Where I part ways is the mechanism. Wanting a slowdown is easy. Engineering one that survives China's incentives, a capability metric nobody can verify, and a deadline that accelerates the exact behavior it's meant to stop is a much harder problem. Plan A only solves the easy one.

I've written before about what happens if we never reach AGI — the slower, stranger transformation nobody puts on a roadmap. This is the mirror image of that argument: the fast version isn't getting governed into a pause either.

Bottom line

Plan A is a strong essay and a weak prediction, and to their credit, the authors say as much: a recommendation, not a forecast, right there on the site. Somewhere between that page and every take treating a 2040 pause as a plausible outcome, the caveat got lost.

It shouldn't have. Nobody with the actual leverage to slow this down has agreed to, and every incentive on the table right now points the other way. Plan for the race. Not the pause.